Samsung 960 Pro NVMe SSD speeds are crazy

So I installed a new Samsung 960 Pro NVMe SSD in my Asus Prime x370 Pro motherboard

I was on the Windows 10 Insider Preview. It was ‘ok’. Used the Samsung Magician software, followed all the rules, rah rah.

samsung mag

Now, Samsung publishes the speed rating for that drive is

Sequential Read Speed:  Max 3,500 MB/sec

Sequential Write Speed:  Max 2,100 MB/sec

 

I mean, what I had was fast. So I didn’t think anything of it. Right?

I had issues with the Insider Preview and rebuilt from scratch. Didn’t have time to debug it, and since I am no longer a Microsoft employee, I don’t have access to private symbols anyway, nor the ability to file a bug. So meh.

Back to 10 Pro. I rebuilt. Before I installed Samsung Magician and the “right” drivers, I did a speed test using Crystal Disk Mark. (Yes, I know it isn’t a diskspd test, I just wanted a quick check).

The results…using stock Windows 10 Pro drivers from 1703 are, shall we say, impressive.

vroom

Now I could install Samsung Magician and retest. But then it’d modify the driver and screw this up for me most likely. So, no thanks. I’ll keep my advertised speeds thank you very much.

Good job to Microsoft, hats off for a stock driver that kicks butt!

default driver

Till next time, the dude does in fact abide, albeit somewhat speedily…

Server 2012 and balanced power plan, part deux – processor queue length

Ok in our last installment on Windows Server 2012 R1 and the balanced power plan, we saw core parking. But was that ‘bad’?

I think so. It depends, naturally, on a variety of factors, but generally speaking in server land, we want stuff done quickly, not don’t slower to save power. Generally.

So here’s what this looks like at high performance plan, running a Monero cryptonote mining program on 4 cores. In the image below you can see we have about 142 hashes / second from 4 cores of work. CPU is at about 62% across all cores, pretty evenly I think.

11

 

So great. What happens if we flip this to balanced plan?

12

We just lost 25% of our cores. 2 parked out of 8. Now, you’ll be quick to note, “hey its teh same CPU util, just busier on 6 cores and idle on 2, so who cares?”

Lets dig a little deeper. Looking at System\Processor Queue Length in perfmon:

13

Ok so we’re bouncing around. Waiting. We have items waiting on a CPU here some of the time. It’s not a lot. But this is a synthetic setup on a VM on a Dell T5500. Not a production server running, say, a stock exchange. So we have a few threads waiting for a core. So what?

Look what happens if we just pop back over to high.14

As soon as I flip to high performance, queue length goes to 0. Because 2 of our 8 cores aren’t parked. And this is in a VM mind you. So there goes the theory that Virtual Machines don’t need a power plan set because they are virtual (I run into this argument from time to time…)

Next installment, we’ll examine the impact to the network stack.

Build a Perfmon data collector set from a PAL xml template

Howdy, simple walk-through post on how to build a Perfmon data collector set using the excellent templates available from Clint Huffman’s PAL utility.

Step 1: Export the appropriate template from PAL;

perfxml

One just clicks Threshold File (1), then Export to Perfmon template file (2), and gives it a file name to save as.

Step 2: Open Perfmon. Expand Data Collector Sets, User Defined, right-click, new – Data Collector Set. Give it a friendly name and hit next.

Step 3: click Browse, pick your XML file created in step 1. Click finish.

Step 4: Right click the new data collector set, properties. click on Stop Condition. Make it as so (200 MB limited file creation, restart at limit). Then click OK.

collectdata

Step 5: right click the PAL_System_Overview under your data collector set, properties, change sample interval to 5 seconds. Click File, set File name format to ddHHmmss and check Prefix file with computer name.

collectdata2

Click OK. right click and start the data collector set. Verify the files are being created in the Perflogs directory.

 

Server 2012 and balanced power plan

Power is something I’ve written about before. I get it. Really I do. I recycle. I like our planet. I want it to be livable for my grand-kids, etc.

But, when work is going to be done on a computer, why you want it to take 2x as long to do the same work, that will end up burning basically the same amount of energy (just less/minute), plus burning my time sitting around waiting on 145 Windows updates to install for hours… is beyond me.

For some troubleshooting I was doing, I needed to validate some things in Windows Server 2012. So I installed it. It took forever. Even on 8 cores, 8GB of RAM, and backed by a solid state drive. Like, an hour to install. And 2 hours to apply 145 updates from Microsoft and it still wasn’t fully done, just the first pass…

So. This is what my system looked like running Windows Update on a Server 2012 (non-R2) vm:

parked

So, CPU 3 is parked. I don’t care much for core parking. Especially not in older OSes on older hardware. Also see how cores 0/1/2 are busy, really busy, and 3/4/5/6/7 are all parked/quiet? Is that good? Why did I allocate 8 cores to this box? Because I want it to get stuff done faster. Not to save power with core parking.

What’s the big deal, you may ask… Here’s the big deal:

Slow Performance on Windows Server when using the “Balanced” Power Plan

But dude, you’ll say, chill out. No. Here’s why:

Hosts Become Unresponsive with XenServer on Nehalem and Westmere CPUs

I quote:

“In addition to the total system lockups, C-states have been linked to a number of less severe issues, such as erratic network performance, resets on storage adapters, random crashed processes, and the overall system performance.”

BIOS to high perf, power plan to high perf, C-states off.

It occurs to me a WU snap in time of what perf looks like isn’t an empirical test, especially for the incredulous reader. So I’ll show you what I mean with a Monero CPU miner shortly…

Part II

Hyper-V Pro Tip: Don’t cross the streams

update: using a single GPU, same results. It’s not the dual-gpu unsupported problem. Something else is hosing my hyper-v host from the guest install.

The results, well, sucked. Nuked the system totally. So don’t cross the streams!

barf

In my lab I was building an MDT share so I could mass deploy endpoints. I’m a hyper-v guy, and a gamer, and an Ethereum miner, so I have a couple video cards in my rig. Now I know, multiple video cards aren’t supported in Hyper-V RemoteFX. I get it. I don’t know why, sounds like some code laziness going on but whatever. I don’t use RemoteFX acceleration in my guest VMs, so I didn’t think much of it…

Until I stood up a Windows 7 Professional SP1 updated install and injected the Hyper-V additions into it as part of the normal MDT process. 

Apparently the guest driver initialization hosed the host? I didn’t think that was supposed to be possible. I’d file a bug but I’m a pro customer now, not enterprise level, so I can’t.

How to find this? Well, GFL looking in the SYSTEM event log, where crashes are typically recorded. It’s empty.

The esoteric Microsoft-Windows-Hyper-V-VMMS/Admin log though? It was pretty helpful. Just have to think to look down in there…

“The required GPU resources could not be accessed. This server cannot run as a RemoteFX host without a GPU. Verify that the GPU is correctly installed.”

gpuerror1

Followed quickly by “The machine contains non-identical GPUs. This is not a supported configuration for RemoteFX.”

gpuerror2

So I get it. “You shouldn’t be running in this config”. Fair enough.

Should Hyper-V allow this configurations’ guest to crash the host? Probably not.

Another fun fact. I don’t use the GPUs for the guest anyway…

gfl

Tanium is hiring!

Hello friends, countrymen, and esteemed guests!

As you may know, I’m a Technical Account Manager* with Tanium now, having left MSFT a bit ago (again). Well, Tanium is growing in big ways and we’re looking for some talented IT/Security folks, particularly in our Southeast Region, and specifically around Tampa, FL and Memphis, TN. We also are on the lookout for folks in Indianapolis, IN and Atlanta, GA!

If you’d be willing to have a conversation about Tanium and careers and seeing if you’d be a good fit here I’d be happy to make time to talk with you as your schedule allows. It’s becoming more and more difficult to find folks with in-depth skills these days so I’m sort of sending out the clarion call here.

The position is a remote/work from home position and encompasses many facets of technology; architecture, security, troubleshooting, automation, consulting, development, and more. This is definitely a job for those that like to stay busy and be challenged every day.

Tanium is a fast moving organization that challenges all of our employees to be our best for both Tanium and our customers. Our customers include a large portion of the US government, 10 of the top 10 banks, most if not all of the major retailers in the US and more.

So hey, hit me up if you want to chat about this opportunity.

jeffrey.stokes@tanium.com

*TAM at Tanium does not mean TAM at Microsoft. It’s sort of a PFE/MCS/TSP role. A floor wax and a dessert topping if you will. It’s fun, fast paced, technical and requires some people skills too.

Misadventures in BIOS updates

So I’m on a new-ish motherboard from ASUS that has an AM4 socket for my Ryzen 1700x processor. So far so good right? Well, ASUS released a new BIOS update that “Update AGESA to 1.0.0.6a”. This sounded awesome, or at least a fix, as BIOS 0803 updated to 1.0.0.6 and was noticeably no longer on the download page…

dudewheresmybios

So I updated. What could possibly go wrong?

So, the new bios 805 appears to have an issue with mixed DRAM configurations where it marks RAM bank B as present, but Windows sees it as hardware reserved and BIOS doesn’t account for it…

IMG_A

So above is the problem case, note memory is 16384, even though DRAM status sees all 4 sticks and they amount to 24GB…

vurt.png

It’s been my experience that computers are generally good at math, at least better than I am at it…so…yeah.

I dorked around, on a whim. I reboot, force D.O.C.P. config (whatever that is). We are now good after 1 reboot for D.O.C.P. to take effect. #becausesunspots

IMG_B.jpg

Yay!

Now, the bad part, when I do this it reboots (naturally), but then fails to post 3 times and recovers). I have to hit F1 (apparently D.O.C.P. means “OVERCLOCKME BADLY”. Since, well, it thinks my DDR4 RAM is capable of 2667 when it is capped at 2400. Whatever.

So I disable D.O.C.P., now I have my memory accounted for properly?!

IMG_C.jpg

Shrug.

Homophobic space nazis here on earth to buy remaindered furniture.

or something.